SAVE-IT
A Final Report of
SAfety VEhicles using adaptive Interface Technology (Phase II: Task 3C):
Driving Performance
(How Do Distracted and Normal Driving Differ:
An Analysis of the ACAS Naturalistic Driving Data)
Prepared by
Paul E. Green University of Michigan Transportation Research Institute,
+1 734 657 0578,
Takahiro Wada University of Kagawa, +81 87 864 2336,
Jessica Oberholzer formerly University of Michigan Transportation Research Institute
Paul A. Green University of Michigan Transportation Research Institute,
+1 734 763 3795,
Jason Schweitzer University of Michigan Transportation Research Institute,
+1 734 730 1396,
Hong Jun Eoh Korea Telecom Research Center,
August 2008
1
TABLE OF CONTENTS
3.0. EXECUTIVE SUMMARY......
Background
Issues......
Method......
Results......
Conclusions......
3.1.PROGRAM OVERVIEW
3.2.INTRODUCTION
Research from SAVE-IT Phase 1......
Other Key Studies......
3.3.Method
Database Examined......
How the Face Clips Were Sampled and Coded......
3.4.RESULTS
1. What are the values of descriptive statistics (e.g., mean, standard deviation, etc.) for common driving performance measures (steering wheel angle, heading angle, throttle opening, and speed)?
2.How do road type and driver age affect those statistics?......
3.How does distraction (as determined by head position) affect those statistics?...
4.What distributions fit those statistics?......
5. For all road types and driver age groups, which single throttle hold definition (sampling interval and size of change threshold (maximum minus minimum)) best distinguishes between normal and distracted driving?
6.As a function of road type, driver age group, driver sex, and how a throttle hold is defined, what are the odds of distracted driving?
7.For each specific road type and driver age group, which throttle hold definition best distinguishes between normal and distracted driving?
8. In addition to throttle holds, which statistics (e.g., mean, frequency above or below some extreme) for which driving-related measures (e.g., lead vehicle range, lane width, outside temperature, etc.) best distinguish between normal and distracted driving?
3.5.CONCLUSIONS
1. What are the values of descriptive statistics (e.g., mean, standard deviation, etc.) for common driving performance measures (steering wheel angle, heading angle, throttle opening and speed)?
2.How do road type and driver age affect those statistics?
3.How does distraction (as determined by head position) affect those statistics?
4.What distributions fit those statistics?
5.For all road types and driver age groups, which single throttle hold definition (sampling interval and size of change threshold (maximum minus minimum)) best distinguishes between normal and distracted driving?
6.As a function of road type, driver age group, driver sex, and how a throttle hold is defined, what are the odds of distracted driving?
7. For each specific road type and driver age group, which throttle hold definition best distinguishes between normal and distracted driving?
8.In addition to throttle holds, which statistics (mean, frequency above or below some extreme value, etc.) for which driving-related measures (lead vehicle range, lane width, outside temperature, etc.) best distinguish between normal and distracted driving?
Considerations for the Future......
3.6.REFERENCES
3.7.Appendix A: Observed Frequency Data
3.8.Appendix B: Descriptive statistics by road superclass, age group and distraction
3.9.Appendix C: additional results of Standard Deviation Change Ratio analysis
3.10.appendix D: comparison of descriptive statistics from distribution and fitted results
3.11.Appendix E: Ratio of Throttle Holds (Hold/NoNhold) by Road superclass, Age group, and distraction
3.12.Appendix F: predictor driving variables
1
3.0. EXECUTIVE SUMMARY
Background
The first objective of Task 3c (performance) was to characterize normal and distracted driving by providing descriptive statistics (e.g., means, standard deviations) and distributions for a wide range of driving performance measures commonly reported in the literature (such as lane position, steering wheel angle, etc.) as well as statistics examined in the prior phase of the project (throttle holds).
The second objective was to determine those statistics that were most effective in distinguishing between normal and distracted driving. Thus, the larger intent is to infer from driving performance in real time when the driver appears to be distracted, and then regulate secondary tasks (by locking them out, warning the driver, etc.) based on that information.
In Phase 1, Task 3a, a summary of the literature on the statistical differences between normal and distracted driving for a wide range of measures was completed (Green, Cullinane, Zylstra, and Smith, 2004). The authors examined 9 well-known papers relating to factors that affect or are affected by driver performance (e.g., SD of steering wheel angle, headway, etc.). Very few dependent measure statistics were consistently reported across studies except for the standard deviation of lane position (8 of the 9 studies, typically 0.25 m with substantial variation). Relationships were developed between that measure and driver age, test conditions, etc. Also noteworthy was the large difference in the standard deviation of velocity between distracted and normal driving (36% on average).
In Task 3b, data was collected for dialing, radio tuning, and destination entry while driving an instrumented vehicle (Zylstra, Tsimhoni, Green, and Mayer, 2003). One of the findings was that during periods of distraction, there was evidence of throttle-holds. During normal driving, there are a large number of small corrections in throttle position that occur, sometimes several per second. When distracted, those corrections may not occur, and speed control becomes more intermittent—that is making corrections, then not making them for a few seconds, and then making them.
Issues
Given the project objectives, prior research, and project needs, the following issues were examined:
1. What are the values of descriptive statistics (e.g., mean, standard deviation, etc.) for common driving performance measures (steering wheel angle, heading angle, throttle opening, and speed)?
2. How do road type and driver age affect those statistics?
3. How does distraction (as determined by head position) affect those statistics?
4. What distributions fit those statistics?
5. For all road types and driver age groups, which single throttle-hold definition (sampling interval and size of change threshold (maximum minus minimum)) best distinguishes between normal and distracted driving?
6. As a function of road type, driver age group, driver sex, and how a throttle-hold is defined, what are the odds of distracted driving?
7. For each specific road type and driver age group, which throttle hold definition best distinguishes between normal and distracted driving?
8. In addition to throttle-holds, which statistics (mean, frequency above or below some extreme value, etc.) for which driving-related measures (lead vehicle range, lane width, outside temperature, etc.) best distinguish between normal and distracted driving?
Method
The contract plan called for re-examining the driving performance data from the ACAS field operational test (Ervin, Sayer, LeBlanc, Bogard, Mefford, Hagan, Bareket, and Winkler, 2005). That naturalistic driving study involved 96 drivers and 136,792 miles of driving. Equal numbers of men and women, in their 20s, 40s, and 60s, participated in the study. There were 9 road types driven: (0)ramp, (1) interstate, (2) freeway, (3) arterial, (4) minor arterial, (5) collector, (6) local, (7) unpaved, and (8) unknown, with many of the analyses examining 3 superclasses of roads (a) interstates and expressways, (b) major roads (arterials and minor arterials), and (c) minor roads (collectors and local roads).
Only data from the baseline condition, when the forward collision avoidance and adaptive cruise control systems were disabled, were examined. Included in the data set are forward scene and face camera views, and data from 400 engineering variables (collected in real time at 10 Hz). The forward scene was recorded at 1 Hz. Distracted driving was determined using the 4 s face video clips collected at 5 Hz, so segments examined consisted of a maximum of 20 data points, sometimes 1 or 2 less depending on system synchronization. Face video clips were recorded once every 5 minutes.
Often, exactly where the driver was looking could not be determined by looking at a face clip. However, as has been the case in prior studies, there was a high correlation between where the driver was looking and head orientation. Therefore, the rule of thumb was that the driver was considered to be distracted when 4 or more consecutive frames occurred where the driver’s head was not oriented toward the forward scene (based on manual coding of at least 2 observers of the 2,914 clips). Given a typical road scene, a glance was about 0.5 s, and some time was required to transition between the road scene and elsewhere. The 4-frame rule provided confidence that frames identified as distracted were truly such. Note that where drivers were looking and whether they were performing a secondary task before or after the 4 s face clip could not be observed.
Results
The analyses produced a large number of figures, tables, and statistics. Only a sample is provided here to give the reader an overview of the extensive results. Tables 3.1 and 3.2 show descriptive statistics for several of the driving performance measures of interest. Many of the distributions were lognormal. As shown in Table 3.3, there were differences in the means of measures related to speed between road super-classes, which makes sense (since the speed limits were different), but also statistically significant differences in the variability of steering input.
Table 3.1. Example Descriptive Statistics
Q1. Histograms and descriptive statistics of overall data for each measureTable 3.2. Histograms as a Function of Driver Age
Q2. Histograms and descriptive statistics of measures by road type & driver ageTable 3.3. ANOVA Results: Effects of Road Superclass and Driver Age
Q2. Do road type & driver age affect driving performance statistics?* (p<0.05), ** (p<0.01), *** (p<0.001), - (no statistical significance)
Driving Performance Measure / Mean / SD
Road Superclass / Age Group / Rd x Age / Road Superclass / Age Group / Rd x Age
Steering Wheel Angle / NA / NA / NA / *** / - / -
Heading Angle / NA / NA / NA / ** / - / -
Throttle Opening / *** / *** / *** / - / - / -
Speed / *** / ** / *** / - / - / -
In terms of overall driving performance statistics, distraction had almost no effect on individual driving performance statistics (Tables 3.4 and 3.5), except for decreasing mean throttle opening by 36% and mean speed by 6%. However, more sophisticated analyses were able to identify differences due to distraction in individual cells of the data set (particular driver age-road type combinations). This led to an effort to fit the distributions of steering wheel angle, heading, and speed (all double exponential) and throttle opening (gamma) for each road type by driver age combination (Table 3.6). Again, at the overall level, there were no differences between normal and distracted driving.
Table 3.4. Effect of Distraction on Driving Performance Statistics
Q3. What are typical values for measures of driving performance?Table 3.5. ANOVA of Distraction and Other Factors on Driving Performance
Q3. Does distraction significantly affect measures of driving performance?* (p<0.05), ** (p<0.01), *** (p<0.001), - (no statistical significance)
Mean
Performance Measure / Road / Age / Dist / Rd x Age / Rd x Dist / Age x Dist
Steering Wheel Angle / - / - / - / *** / - / **
Heading Angle / *** / *** / ** / *** / *** / *
Throttle Opening / *** / *** / * / *** / - / -
Speed / *** / *** / - / *** / *** / -
SD
Performance Measure / Road / Age / Dist / Rd x Age / Rd x Dist / Age x Dist
Steering Wheel Angle / *** / - / * / - / - / *
Heading Angle / ** / - / - / - / - / -
Throttle Opening / - / - / - / - / - / -
Speed / - / - / ** / - / * / -
Table 3.6. Fit of Distributions to Various Road Type-Driver Age Combinations
Q4. Fit model to input and output measures of interestSteering Wheel Throttle / Limited access road, middle age drivers
Mean Fit: Good in general
SD Fit:
Steer Error = 10-50% (fit with SD)
Heading Error=3-50% (fit with SD)
Throttle Error = 1-12%
Speed Normal Error = 1-20%
Distracted Error = 3-50%
Given the results of Phase 1, throttle-holds were given special attention. Throttle-holds are defined by 2 parameters, the time window over which it occurs and the threshold (percentage change) required. As shown in Table 3.8, differences between distracted and normal driving were most likely to occur for 1-s time windows, and the statistic was most sensitive for limited access roads for middle-age drivers, the most stable combination. What makes no sense at all is why throttle-holds were more likely for normal driving, the direct opposite of what the theory predicts. Given the 4-s limit of the data, the true intermittent nature of throttle-holds could not be explored with this data set.
Table 3.7. Throttle-Hold Parameter Estimation
Q5. Comparison of various throttle hold parameters by road type and driver ageEffect of changing parameters (Limited access road, young drivers)
/ Highest throttle hold frequency with smaller time window & larger threshold
Most consistent throttle holds parameters for all road x age combinations when:
Time window = 1 sec
Threshold = 4
(works best for limited access roads & for middle age drivers)
However, as with the prior analyses, the most sensitive detection of normal vs. distracted differences occurred when the statistic of interest, here throttle-holds, was tailored to the road class-driver age combination of interest, with larger thresholds being best for minor roads. Table 3.8 shows the best parameters for each case.
Table 3.8. Best Throttle Hold Parameters for Each Road Class-Driver Age Combination
Q7. Effectiveness of using throttle holds to identify distracted drivingusing road type and age specific parameters
Comparison of throttle hold parameters (Major roads, older drivers) /
Best throttle hold parameters
for each road type, age combination
In the final major analysis, logistic regression identified other statistics and factors that discriminated between normal and distracted driving. For limited access roads (7.6% distraction in baseline), the 6 best predictors for distracted driving were:
- Turn signal [(0-baseline) off, (1) on]
- Age group [(0) young, (2) middle, (3) old]
- Speed from transmission –m/s [(0) .05<=x<=.95, (1) x<.05 or x>.95]
- Velocity of current in-path vehicle – m/s [(0) x=0, (2) 0<x<=30, (3) x>30
- Deceleration of current in-path vehicle - percentile [(0) x>=.05, (1) x<.05]
- Lane offset confidence [(0) none, (2) low/medium, (3) high]
For major roads (4.4% distraction in baseline), the 6 best predictors for distracted driving were:
- Gender [(0) female, (1) male]
- Range to closest in-path vehicle [(0) x=0, (2) 0<x<=60, (3) x>60]
- Forward road geometry – 40m [continuous]
- Brake [(0) not active, (1) active]
- Lane offset confidence [(0) none, (2) low/medium, (3) high]
- Peak to peak vertical acceleration from ABS – g [continuous]
For minor roads (1.3% distraction in baseline), the 6 best predictors for distracted driving were:
- Lane width – m [continuous]
- Outside temperature – C [(0-baseline) .05<=x<=.95, (1) x<.05 or x>.95]
- Speed from transmission – m/s [continuous]
- Forward road geometry – 120m [continuous]
- Heading offset from lane center – m [(0) .05<=x<=.95, (1) x<.05 or x>.95]
- Deceleration of current in-path vehicle [continuous]
Conclusions
This report contains a wealth of statistics and figures that describe normal and distracted driving as a function of the type of road driven and driver age and sex. Data provided includes the mean and standard deviation, as well as the distribution type, for each combination, all of which should be extremely useful for modeling of driver performance.
Particularly striking was the lack of significant differences due to distraction when driver performance statistics such as the overall mean speed were examined (though the standard deviation of speed was significantly different).
In part, this is because the effects of distraction are complex, and differences become more apparent when particular road type-driver age (and sex) combinations are considered. This suggests the need for a more sophisticated approach than brute force analysis, which will make real-time determination of distraction quite difficult. Overall, key driving task related parameters were turn signal use for expressways, lead vehicle range (when greater than 60m) for major roads, and lane width, lane offset, and lead vehicle velocity for minor roads.
Further, analysis of throttle-holds showed mixed results. The optimal combination was a 4% threshold for a 1-s window, but the overall effects were in the opposite direction as expected. Throttle-hold measures were much more sensitive when tailored to the road class-driver age combination of interest.
1
3.1.PROGRAM OVERVIEW
Driver distraction is a major contributing factor to automobile crashes. The National Highway Traffic Safety Administration (NHTSA) has estimated that approximately 25% of crashes are attributed to driver distraction and inattention (Wang, Knipling, & Goodman, 1996). Recent estimates from the 100-Car study suggest that distraction may contribute to more than three quarters of all crashes (Dingus, Klauer, Neale, Petersen, Lee, Sudweeks, Perez, Hankey, Ramsey, Gupta, Bucjer, Doersaph, Jermeland, & Knipling, 2006). The issue of driver distraction may become more critical in the coming years because increasingly elaborate electronic devices (e.g., cell phones, navigation systems, wireless Internet and email devices) are being brought into vehicles that may further compromise safety. In response to this situation, the John A. Volpe National Transportation Systems Center (VNTSC), in support of NHTSA's Office of Vehicle Safety Research, awarded a contract to a diverse team led by Delphi Electronics & Safety including Ford, the University of Michigan Transportation Research Institute (UMTRI) and the University of Iowa. The goal of this program was to develop, demonstrate, and evaluate the potential safety benefits of adaptive interface technologies that manage the information from in-vehicle systems based on real-time monitoring of the roadway and the state of the driver. The contract, known as SAfety VEhicle(s) using adaptive Interface Technology (SAVE-IT), is designed to mitigate distraction with effective countermeasures and enhance the effectiveness of safety warning systems.